Darrell Conklin and Mathieu Bergeron
Department of Computing
City University London
United Kingdom
conklin@city.ac.uk
Pattern discovery is an important part of computa-
tional music-processing systems. The discovery of
patterns repeated within a single piece is an impor-
tant step to segmentation according to thematic
structures (Ruwet 1966). Patterns found within a
few works may be signatures that can be instanti-
ated for style emulation of novel musical material
(Cope 1991; Rowe 1993) and can reveal a deep
similarity in musical material. Patterns that are
conserved across many pieces in a large corpus can
represent structural building blocks and used for
comparative style analysis and music genre recogni-
tion (Huron 2001; Conklin and Anagnostopoulou
2001; Lin et al. 2004).
Pattern discovery methods can be discussed
according to the expressiveness of patternsin
particular, the levels of abstraction permitted by
pattern components. Many approaches are re-
stricted to a representation in which every pattern
component is described using the same musical
attribute: pitch, duration, interval, or fixed combi-
nations of these (e.g., linked interval / duration, etc.).
In these approaches, an event has only one possible
representation, and therefore patterns can be effi-
ciently found using general string algorithms (Gus-
field 1997) after transforming the corpus to strings
of attribute values.
Recent methods have considered whether this
restriction can be relaxed by allowing patterns with
heterogeneous components and subsumption rela-
tions among possible pattern components (Lartillot
2004; Cambouropoulos et al. 2005; Conklin and
Bergeron 2007). The need for such patterns can be
motivated with a few melodic fragments (see Figure
1) from the music of the famous twentieth-century
French singer and songwriter Georges Brassens
(19211981). In both pairs of fragments, the descrip-
tion of events by melodic interval or melodic con-
tour alone is inadequate. Though the fragments
within each pair have a common duration pattern,
Computer Music Journal, 32:1, pp. 6070, Spring 2008
2008 Massachusetts Institute of Technology.
60
Feature Set Patterns in
Music
there is no melodic interval pattern that spans the
complete fragments, though some events do have
conserved melodic intervals.
This article describes a new approach to pattern
representation and discovery in music, where pat-
tern components can contain any number of fea-
tures and can be as general or as specific as required
by the data. An important concept in this work is
subsumption, which provides a natural way to
explore the pattern search space in a general to
specific manner, pruning entire branches when
patterns become infrequent. The space of patterns
that may be discovered is very rich, and patterns
have highly flexible levels of abstraction, much
more than is possible with single attribute patterns.
A well-known fact of knowledge representation is
that increased expressiveness usually leads to in-
creased time complexity of reasoning (Brachman
and Levesque 2004). For pattern discovery in par-
ticular, allowing pattern components to contain a
flexible and varying number of attributes substan-
tially increases the size of the pattern search space.
Though the size of this space is substantially re-
duced by the restriction to frequent patterns, when
there are too many frequent patterns in the search
space, heuristic algorithms must be employed.
This article presents two algorithms to solve these
pattern discovery problems: a complete algorithm
and a heuristic probabilistic algorithm.
A large number of patterns can be revealed in a
single piece or corpus, and these patterns must
somehow be filtered and ranked for presentation.
In this article, a statistical measure is used to order
patterns according to the deviation of their observed
from their expected frequency in the corpus.
The methods developed here are applied to the
music of Georges Brassens. The style of Brassens,
described by Keefe (2002) as sophisticated simplic-
ity, is well suited for pattern discovery studies. It
did not change substantially throughout his career,
and his output was prolific: he set over 170 texts to
music, in addition to over a dozen texts of other
well-known French poets.
Computer Music Journal
Figure 1. (a)(b): First two
phrases of Brassenss Le
Bistrot; (c)(d): end of the
first two phrases of
Brassenss Tonton Nestor.

(a) (b)

(c) (d)

Methods Table 1. Glossary of Terminology and Notation

Used in This Study
This section provides the necessary definitions and
Term Definition
an overview of the algorithms for pattern discovery.
Some of the definitions and terminology are stan- viewpoint a function computing a property of a
dard in the data mining literature (Agrawal and music event
Srikant 1995; Ayres et al. 2002; Uno et al. 2003), but feature a viewpoint / value pair
there are also a few novelties dealing specifically
with music data mining. For reference, Table 1 an undefined value
provides a summary of notation and terminology feature set a set of features representing a
used in this section. logical conjunction
{} the empty feature set

Events and Viewpoints closed feature set no subsumee feature set has the
same total count in the corpus
An event is a music object within a sequence. feature set pattern a sequence of feature sets occurring
Music objects can be notes, or structured objects in the corpus
such as simultaneities (Conklin 2002) and se-
Cp() the number of pieces having at least
quences (Conklin and Anagnostopoulou 2005). In
one occurrence of the specified
this article, we focus on melody, and we only con- feature set or pattern
sider note objects. A viewpoint is a function that
computes values for events in a sequence (Conklin Ct() the number of non-overlapping
and Witten 1995; Conklin 2006). For example, for occurrences of a feature set or a
the melodic interval viewpoint, the values in the pattern in the corpus
range of the viewpoint are integers representing the Et() pattern expected total count
difference in semitones between an event and its
I() pattern interest measure
predecessor.
Table 2 provides a catalog of viewpoints used in frequent occurring in the corpus with at least
this study. At the top of the table are the primitive the specified piece and total count
viewpoints, namely, those used to select the basic thresholds
attributes of notes. Following this are derived view- infrequent occurring in the corpus below the
points that compute values using the primitive specified piece count or total count
viewpoints. They can be further grouped into unary threshold
(computable from a single event, e.g., pitch class) and maximal frequent a more specific frequent
binary (requiring one preceding event to compute a pattern pattern cannot be found in the
value, e.g., melodic interval or melodic contour). A corpus
potentially unbounded set of derived viewpoints is
available through the use of constructors, which

Conklin and Bergeron 61

Table 2. Viewpoints Used in This Study build new viewpoints from existing ones. In this
study, constructors are not employed, but rather a
Viewpoint Description Range Set
fixed set of derived viewpoints is used.
pitch pitch of event as MIDI Viewpoints are used to transform sequences of
number {0, . . . , 127} concrete events into more abstract sequences. This
key MIDI key signature as is done simply by applying a viewpoint to every
number of sharps (+) event in a sequence. Table 3 illustrates how events
or flats () {7, . . . , +7} in a short melodic fragment (at a resolution of 192
onset onset of event in
ticks per beat) are transformed using the viewpoints
MIDI ticks Z+ of Table 2. The fragment is notated using three flats
in the MIDI key signature (key:-3). Though this
duration duration of event in fragment (and the MIDI files for the Brassens corpus
MIDI ticks Z+
used in the current study) does not contain indica-
ml1 first metric level {t} tion of mode or spelling of pitches, some quasi-
ml2 first or second metric diatonic features can still be computed from the
level {t} MIDI key signature. To do so, major mode is as-
sumed (e.g., E-flat major or ref:3), and quasi-
pc pitch class {0, . . . , 11}
diatonic intervals are computed between pitches
ref pitch class of tonic, standardized to that major key (whole scale steps
assuming major mode {0, . . . , 11} receive a natural n and modified scale steps an
intref pitch class interval augmented s sign on their interval).
from ref {0, . . . , 11} Table 3 also illustrates how a hierarchical attri-
int pitch interval Z bute is represented using feature sets. An event can
occur in the highest metric level beginning a bar
intu unordered pitch interval N (ml1:t and ml2:t), only in a second level (ml2:t;
intpc pitch class interval {0, . . . , 11} in triple meter, both the second and third beats are
intupc unordered pitch class in this level), or in neither (). Note that any binary
interval {0, . . . , 6} viewpoint is undefined () for the first event in the
fragment, and that the intscs and inuscs view-
intscs scale steps; natural or
points are defined only if the first pitch is a major
raised in major mode Z {n, s}
scale degree. Some viewpoints (contour classes,
intuscs unordered scale steps N {n, s} metric level viewpoints) return only a possible
repeat contour class {t} value of t (true). Negated classes for these view-
points (e.g., not a step, not in the highest metric
up contour class {t}
level) can easily be handled, but they do not usu-
down contour class {t} ally represent musically significant pattern compo-
step contour class {t} nents and are not used in this study.
leap contour class {t}
ioi inter-onset interval in Feature Sets and Patterns
MIDI ticks Z+
dc3 three-point duration A viewpoint / value pair is called a feature. These are
contour {<,=,>} notated by a viewpoint name and value separated by
dr duration ratio Q+ a colon, for example, pitch:63 or int:-4. A fea-
ture set represents a logical conjunction of features.
Here, N denotes natural numbers (including 0), Z all integers, Z+ A feature set is instantiated by an event if the event
positive integers, Q+ positive rational numbers, and t true.
has every feature within the set; the empty feature

62 Computer Music Journal

Table 3. Fragment of Brassenss Une jolie fleur Transformed Using the Viewpoints of Table 2

Primitive viewpoints
pitch 63 66 67 63 66 67 63 66 67 65
key 3 3 3 3 3 3 3 3 3 3
onset 0 192 288 384 576 672 768 960 1152 1344
duration 192 96 96 192 96 96 192 192 192 192
ml1 t t
ml2 t t t t
Derived viewpoints (unary)
pc 3 6 7 3 6 7 3 6 7 5
ref 3 3 3 3 3 3 3 3 3 3
intref 0 3 4 0 3 4 0 3 4 2
Derived viewpoints (binary)
int +3 +1 4 +3 +1 4 +3 +1 2
intu 3 1 4 3 1 4 3 1 2
intpc 3 1 8 3 1 8 3 1 10
intupc 3 1 4 3 1 4 3 1 2
intscs 1s 2n 1s 2n 1s 1n
intuscs 1s 2n 1s 2n 1s 1n
repeat
up t t t t t t
down t t t
step t t t t
leap t t t t t
ioi 192 96 96 192 96 96 192 192 192
dc3 < = > < = > = = =
dr 1/2 1/1 2/1 1/2 1/1 2/1 1/1 1/1 1/1
The onset time of the fragment has been shifted to time 0.

set {} is instantiated by any event. For example, the
feature set {pitch:63,int:-4} is instantiated by
the events at positions 4 and 7 of the melody frag-
ment in Table 3. A feature set can be specialized by
adding one or more features to the set. The more
general feature set is said to subsume the special-
ized feature set: all instances of the specialized
feature set are also instances of the more general
feature set. For example, the feature set {pitch:63}
subsumes the feature set {pitch:63, int:-4}
which in turn subsumes {pitch:63, int:-4,
ml1:t}.
The subsumption relation between feature sets
can be visualized as a taxonomy in which nodes of
the taxonomy are feature sets, directed links repre-
sent subsumption, transitive subsumption links are
omitted, and a node is placed between its most spe-
cific subsumers and its most general subsumees.
Figure 2 shows a small feature set taxonomy
using the five viewpoints step, down, repeat,
leap, and up. These viewpoints are used to repre-
sent various levels of abstraction of melodic con-
tour classes (Conklin and Bergeron 2007). Note that
this taxonomy is not presented as a full lattice of
subsets, because some feature sets (e.g., the set
{step:t, repeat:t}) cannot be instantiated by
any event and are therefore contradictory.
A pattern is a sequence of feature sets. A pattern

Conklin and Bergeron 63

Figure 2. Taxonomy of
melodic contours repre-
sented as feature sets. The
inverted triangle repre-
sents the empty feature set.
down:t step:t repeat:t leap:t up:t
step:t leap:t step:t leap:t
down:t down:t up:t up:t
occurs in (i.e., is instantiated by) a contiguous event
sequence if the component feature sets of the pat-
tern are instantiated by the corresponding events.
For example, the pattern [{int:+3,dc3:<},
{int:+1}] occurs at positions 2 and 5 of the mel-
ody fragment in Table 3. In a corpus, the total count
of a pattern is the number of positions in which it
occurs (not counting overlapping occurrences). The
piece count of a pattern is the number of pieces that
have one or more occurrences of the pattern. A
frequent pattern is one that occurs with at least a
specified minimum total count and piece count.
Consequently, a frequent pattern will not contain
infrequent feature sets as components. A maximal
frequent pattern is a frequent pattern whose compo-
nent feature sets cannot be further specialized with-
out the pattern becoming infrequent.
Pattern Interest Measure
Patterns can be ranked according to the difference
between observed and expected counts in a corpus
(Conklin and Anagnostopoulou 2001; Huron 2001),
similar to methods for ranking word collocations in
natural language (Manning and Schutze 1999). Large
differences between observed and expected counts
indicate potentially interesting patterns. Here, a
pattern P is given a pattern interest measure I(P)
that represents the ratio of observed to expected
counts:
Ct(P)
I(P) = (1)
Et(P)
64
where Ct() represents the observed total count in a
corpus and Et() the expected total count. To define
the expected total count, consider a feature set
pattern P = f1, . . . , fn, and let N = Ct({}) (i.e., N is the
total number of events in the corpus). The (non-
overlapping) total count of the pattern can be no
more than N / n (the number of events in the corpus
divided by the length of the pattern). The expected
total count in the corpus is this maximum quantity
multiplied by the probability of the pattern:
N
Et(P) = p( f1 , . . . , fn) (2)
n
The probability of the pattern is computed using a
zero-order model (assuming independence of suc-
cessive events); this is the product of the relative
frequencies within the corpus of the component
feature sets of the pattern:
n
Ct( fi)
p( f1 , . . . , fn) = (3)
i =1 N
The pattern interest can be computed rapidly since
the raw feature set counts emerge directly from the
first phase of the pattern discovery algorithm, as
described below.
Pattern Discovery Algorithms
In this section, two algorithms for pattern discovery
are developed. The first algorithm is complete and
finds all maximal patterns, and the second is an
optimization algorithm that uses a probabilistic
hill-climbing approach and is suited to rapidly find a
single interesting pattern in one or two pieces. This
algorithm is heuristic and therefore guarantees nei-
ther that the pattern found will be maximal nor the
most interesting, but it can be used to search large
search spaces that may arise when the specified
minimum total and piece count are low.
Complete Algorithm
A corpus (which may be one, two, or any number of
pieces) is first transformed to feature-set sequences
by saturation, i.e., applying each viewpoint in a
catalog to every event in every piece (as illustrated
Computer Music Journal
Figure 3. Taxonomy of
closed frequent feature
sets in Brassens, piece
count 132.

dc3:>
2921 0.26

intupc:2
3354 0.30

intref:4
1944 0.17

down:t
4635 0.41

intu:2 intu:2
dc3:=
intupc:2 intupc:2
dr:1/1
step:t intuscs:1n
5460 0.48
3294 0.29 step:t
11273 1.00
3114 0.28
dc3:<
intuscs:1n
2760 0.24
step:t
4338 0.38 intuscs:1n
step:t
step:t
4966 0.44
step:t ml2:t intuscs:1n
ml2:t 2264 0.20 step:t
intref:0 2604 0.23 ml1:t
2208 0.20
step:t ml2:t
ml1:t ml1:t 1414 0.13
ml2:t ml2:t ml2:t
5392 0.48 2958 0.26 1614 0.14

up:t up:t
4366 0.39 ml2:t
2132 0.19

in Table 3). The pattern discovery algorithm then
proceeds in two phases. In the first phase, all fre-
quent feature sets are found and configured into a
taxonomy using a description logic classification
algorithm (Brachman and Levesque 2004), which
places each feature set between its most specific
subsumers and its most general subsumees. Because
a frequent pattern cannot have an infrequent feature
set as a component, the potential feature-set space
may be restricted to those feature sets that actually
occur frequently in the corpus. Furthermore, it
suffices to consider only closed feature sets, namely,
those that do not subsume any other feature set
with the same total count. This is because all com-
ponents of a maximal pattern must by implication
be closed feature sets. The set of all closed feature
sets that are frequent in their total count can be
efficiently computed using the method of Uno et al.
(2003). This set is further pruned to those having at
least the minimum specified piece count.
Figure 3 displays the subsumption taxonomy of
closed feature sets from a corpus of 132 pieces com-
prising 11,273 events (the Results section describes
this corpus in more detail), using the viewpoints
provided in Table 2 and restricted to feature sets
occurring in all 132 pieces. At the bottom of each
node of the taxonomy are the total count of the
feature set and the relative frequency of the feature
set in the corpus. For example, 48% of all events
have the feature ml2:t, and 19% of all events are
further specialized to include the feature up:t.
The second phase of the discovery algorithm
explores the specialization space of frequent feature
set patterns in the search for maximal frequent pat-

Conklin and Bergeron 65

terns. This space can be fully visited by recursively
applying two refinement operators (Ayres et al. 2002)
beginning at the empty pattern: an I-step, which
specializes the right-most component of a pattern
by walking one step down the feature set subsump-
tion taxonomy (there may be several such special-
izations), and an S-step, which appends the empty
feature set {} to a pattern. For example, referring to
the closed feature set taxonomy of Figure 3, I-steps
applied to the single component pattern [{step:t}]
would yield the three patterns [{intu:2,
intupc:2,step:t}], [{intuscs:1n,step:t}],
and [{step:t,ml2:t}]. The S-step would yield
the single pattern [{step:t},{}].
A depth-first search using these refinement opera-
tors is used to explore the pattern space. All search
nodes have an instance map, implemented as an
associative array that maps piece / position pairs to a
Boolean value indicating whether the pattern at the
node occurs in the piece at the indicated position.
When an S-step or I-step is applied, a new instance
map is generated by intersection with the maps of
the empty feature set (S-step) or the specialized
feature set (I-step), taking into account an appropri-
ate position offset. The empty feature set {} has a
fully saturated instance map, as it occurs in every
piece at every position. After the initial construc-
tion of the instance maps for all closed feature sets,
the corpus does not need to be probed again, and
pattern counts emerge directly from instance-map
intersection. If the pattern counts indicate that a
candidate specialization is infrequent, the current
branch of the search can be safely terminated. This
is because any pattern that is a specialization of an
infrequent pattern will also be infrequent.
To avoid redundant computation, it is important
to avoid visiting any pattern in the search space
more than once. This can be achieved by converting
the taxonomy into a tree after the first phase by
breaking all but one subsumption link pointing
into a node, thereby ensuring that no feature set is
visited by more than one chain of I-steps. The
dashed links in Figure 3 indicate one possible set of
subsumption links that can be removed to produce
a tree.
An S-step (appending the empty feature set) ap-
plied to a frequent pattern will usually produce only
66
minimal changes to its instance map and will there-
fore usually lead to a frequent pattern. To avoid
reporting maximal patterns that end with chains of
empty feature sets, every node in the search space
contains a pointer back to the node of the last I-step
in the specialization chain, and when a maximal
frequent pattern is reported, this link is followed to
determine the appropriate pattern for presentation.
Avoiding patterns beginning with an empty feature
set is accomplished simply by requiring the first
refinement applied in the search to be an I-step.
Heuristic Probabilistic Algorithm
The size of the search space of frequent patterns is
determined by several factors including the number
of viewpoints in the catalog, the size of the corpus,
and the specified piece-count and total-count
thresholds. For corpus analysis tasks in which
patterns are required to occur in a substantial
fraction of the pieces, entire branches of the search
space can be pruned quite early. For larger search
spaces, a heuristic probabilistic hill-climbing
method can be used. This method is similar to the
Gibbs sampling methods used for protein motif
discovery (Lawrence et al. 1993). The interest mea-
sure (Expression 1) is computed for all candidate
specializations (produced by all possible I-steps and
the S-step), and one of these is sampled. The prob-
ability p of sampling a pattern P from the set Q of
all candidate specializations is
I(P)
p(P,Q) = (4)
q Q I(q)
and the sampled pattern then becomes the new
current pattern. The process begins at the empty
pattern and is iterated until there are no further
candidate specializations (i.e., until no candidate
specialization is frequent). To reduce the effect of
climbing to solutions with low pattern interest, a
random restart method is employed. The probabilis-
tic hill-climbing search is restarted a fixed number
of times, and the most interesting pattern arising
from all iterations is reported. In the heuristic algo-
rithm, the taxonomy of frequent feature sets is not
converted to a tree, because it is necessary to in-
Computer Music Journal
spect all specializations of a pattern to compute the
set Q (used in Expression 4).
Results
Georges Brassenss songs (Brassens 1993) in MIDI
format were collected from various online sources
and prepared for analysis. The onsets and durations
of notes were appropriately quantized, the melody
line was isolated, instrumental introductory and
concluding material were removed, and repetitions
of stanza and chorus material were also removed.
Internal key and time signature changes were cor-
rectly handled. This process led to a corpus of 132
melodies, comprising a total of 11,273 notes. The
corpus was then saturated as described herein with
all 23 viewpoints presented in Table 2.
Three types of analysis are illustrated here using
the pattern discovery method: corpus analysis
(patterns found in many pieces), comparative anal-
ysis (patterns found in two pieces), and intra-opus
analysis (patterns occurring two or more times in a
single piece). Patterns were allowed to have at most
one empty feature set, and they were required to
have a pattern interest of greater than 1. For com-
parative and intra-opus analyses, the piece count
was set to 2 and 1, respectively, the minimum total
count was set to 2, and the heuristic probabilistic
algorithm was used with 10 iterations.
Corpus Analysis
With a piece-count threshold of 132 pieces (i.e.,
requiring patterns to be present in all pieces in the
corpus), there are 20 closed frequent feature sets
(Figure 3) and only three general patterns are re-
ported (the top three patterns of Table 4). With a
piece-count threshold of 125 pieces (i.e., 95% of the
corpus), there are 134 closed frequent feature sets,
and 100 patterns with interest greater than 1 are
found; three selected patterns are displayed in Table
4. The first pattern refers only to duration features,
the second to a mix of duration and pitch features,
and the last only to pitch features. The last pattern
is chosen here to illustrate the results. This pattern
Table 4. Some Feature Set Patterns Found in the
Brassens Corpus
Pattern Cp Ct
[{dc3:>},{dc3:<}] 132 2114
[{down:t},{up:t}] 132 1935
[{down:t},{ml2:t}] 132 1852
[{dc3:>},{dc3:<},{},
{dc3:=, dr:1 / 1}] 129 988
[{step:t},{down:t,
ml2:t},{dc3:<}] 125 602
[{up:t},{},{intuscs:1n,
step:t},{down:t}] 126 760
The top half shows three patterns found in all pieces; the bottom
half contains three patterns found in at least 125 pieces.
has length 4 and covers an event sequence of length
5, because the first feature set contains the binary
feature up:t. Instances of the pattern begin with a
melodic motion upwards, then any event can follow
and the pattern ends with an event having a down-
ward melodic motion. In addition, the penultimate
event in the sequence is further restricted according
to a scale step feature and must be a step of exactly
one scale degree. The pattern occurs in 126 pieces of
the corpus and has a total count of 760.
The musical relevance of the pattern can be il-
lustrated by two instances of the pattern within
the song Le vieux Lon (see Figure 4: the pattern
covers the last five events of each fragment). Note
how the empty feature set in the second position
is filled by a leap up in the first instance and by a
step movement down in the second. Though the
presented pattern is maximal for the corpus, it is
not for the particular instances of Figure 4. The
instances from Le vieux Lon could be captured
more precisely by specialization, for example, add-
ing the feature intscs:-1n to the last feature set.
The resulting specialized pattern, however, would
not be maximal, as it would be found within less
than 95% of all pieces in the corpus. Finally, note
that the third component of the pattern could be
presented more succinctly, because the feature
intuscs:1n implies the feature step:t. (This is
further highlighted by the fact that the feature set
Conklin and Bergeron 67
Figure 4. Two instances of
the last pattern of Table 4
in Brassenss Le vieux
Lon.
intuscs:1n is not closed in the corpus; see
Figure 3.)
The patterns in Table 4 all have an observed to
expected ratio greater than 1, computed using the
zero-order model of expectation (Expressions 2 and
3). To verify this against a different model of expec-
tation, the Essen folksong collection (Schaffrath
1995), comprising approximately 6,000 songs con-
taining a total of about 300,000 events, was scanned
with each pattern in Table 4. All patterns were
found to have higher relative total count in the
Brassens corpus. Interestingly, the fifth pattern in
Table 4 has a relative total count three times that of
the Essen corpus. This pattern covers four events,
with a melodic motion of a step, a motion down to
the second metric level, followed by a note of
shorter duration.
Comparative Analysis
Comparative analysis, taken here to mean the com-
parison of two pieces for shared patterns, is a very
challenging task outside of the typical application
of comparison of melodic variations of a theme.
Figure 5 shows two phrases from Brassenss Les
quatre bacheliers and Rien jeter. The pieces have
a similar overall form: four phrases, each having
initial notes in a convex shape, followed by a few
cadential notes. A pattern was found that aligns the
first two phrases of the pieces; the feature set of one
component of the pattern, covering the B3 / E4 in the
second bar, is
{intuscs:1n, intref:2, down:t, step:t, dc3:>,
dr:3 / 1, ml2:t, duration:288, ioi:96,
int:-2, intu:2, intpc:10, intupc:2,
intscs:-1n}
This pattern component refers to several specific
features: an interval of two semitones from the
68
Figure 5. First two phrases
of Brassenss (a) Les quatre
bacheliers and (b) Rien
jeter.
(a)
(b)
major tonic (intref:2), a melodic interval of 2
occurring at the second metric level, with a dura-
tion of a dotted eighth, three times that of the pre-
ceding note. Note again that this pattern component
is not presented in the most compact form, as the
melodic interval feature int:-2 implies several
other features (e.g., intu:2, step:t, down:t) that
could be omitted to improve readability.
Intra-Opus Analysis
The motivating musical fragments presented earlier
in Figure 1 were both identified using the probabi-
listic heuristic, pattern discovery method. In Le
Bistrot, a pattern of length 15 is discovered, cover-
ing exactly the first two phrases of the piece. For
example, the pattern components aligning the first
and last events of the Le Bistrot fragments are the
feature sets
{pitch:60, pc:0, intref:9, key:-3, ref:3,
ml1:t, ml2:t, duration:96}
{key:-3, ref:3, step:t, dc3:>, ml2:t, ioi:96,
intu:1, intupc:1}
The last feature set is satisfied by any event in-
volving one scale step (either up or down) a key of
3 (major tonic 3, or E-flat major), a longer duration
than previous note, and an inter-onset interval of
one eighth-note in duration. This and all feature
sets appearing in the pattern are naturally quite
specific, as they are required only to occur twice in
the piece.
In Tonton Nestor, a pattern of length 14 is discov-
ered, covering the two fragments presented in Fig-
ure 1. For example, the pattern components aligning
the first and last events of the Tonton Nestor frag-
ments are the feature sets
Computer Music Journal
{intuscs:r, key:4, ref:4, repeat:t, dc3:<,
dr:1 / 3, ml1:t, ml2:t, duration:64,
ioi:192, int:0, intu:0, intpc:0, intupc:0,
intscs:0}
{key:4, ref:4, leap:t, dc3:>, dr:4 / 1, ml1:t,
ml2:t, duration:384, ioi:96}
The last feature set refers to a key of 4 (major
tonic E major), a duration of a half-note, and a leap
to a longer event (with a duration of four times the
previous note) on a strong beat.
Discussion and Future Work
This article has developed and applied a new pattern
representation for music data mining. The represen-
tation, inspired by sequential data mining represen-
tations (Agrawal and Srikant 1995), allows any
number of features in components of patterns, with-
out restriction to a single viewpoint. The feature-set
representation captures and extends recent innova-
tions (Lartillot 2004; Cambouropoulos et al. 2005;
Conklin and Bergeron 2007) in the field of musical
pattern discovery in a simple and elegant way. For
pattern discovery, pattern components may con-
tain as many or as few features (including the pos-
sibility of an empty feature set) as are necessary to
ensure that the whole pattern is frequent yet also
maximal. The methods were illustrated on melo-
dies by the French composer and songwriter
Georges Brassens.
The tasks of melodic similarity (Mongeau and
Sankoff 1990) and intra-opus pattern discovery in
music are closely related; conserved patterns can
indicate melodic similarity. In contrast to the work
on melodic similarity and on sequential pattern
mining (Agrawal and Srikant 1995), the pattern-
discovery method developed here does not permit
insertions or deletions of events while computing
pattern occurrences. Such a facility could be intro-
duced, but care should be taken in the algorithm to
handle binary features that refer to previous events
and to manage the increase in the complexity of
pattern-occurrence counting.
The method used for ranking patterns is a statisti-
cal measure based on the deviation of observed from
expected pattern total count in the corpus. In this
study, for simplicity, the analysis corpus itself was
used to construct a zero-order model of pattern
component frequencies, which, though not optimal,
does in practice yield some interesting patterns
toward the top of the report. More realistic back-
ground models and interest measures employing
relative frequencies in comparison repertoires could
be explored.
The aim of this research is that an analyst should
have available a large feature catalog and should not
need to restrict a priori the features used to describe
the data nor to predict how they will combine into
interesting feature sets and patterns. The algorithm
and results presented here suggest the practicality
of moving towards this more expressive representa-
tion for music patterns.
Acknowledgments
An earlier report of this work was presented at the
IJCAI 2007 Music-AI Workshop (Bergeron and
Conklin 2007), and we thank the workshop organiz-
ers for making the event possible. Taxonomies were
drawn with the Graphviz tool and music layout
performed with the Lilypond package. Mathieu
Bergeron is supported by a scholarship from Fonds
Qubcois de la Recherche sur la Nature et les
Technologies (FQRNT) and a Ph.D. fellowship from
City University London.
References
Agrawal, R., and R. Srikant. 1995. Mining Sequential
Patterns. Proceedings of the Eleventh International
Conference on Data Engineering. New York: IEEE
Computer Society Press, pp. 314.
Ayres, J., et al. 2002. Sequential Pattern Mining Using a
Bitmap Representation. Proceedings of the Interna-
tional Conference on Knowledge Discovery and Data
Mining. New York: ACM Press, pp. 429435.
Bergeron, M., and D. Conklin. 2007. Representation and
Discovery of Feature Set Patterns in Music. Proceed-
ings of the International Workshop on Artificial Intel-
Conklin and Bergeron 69
 ligence and Music. San Francisco, California: Morgan
Kaufmann, pp. 112.
Brachman, R., and H. Levesque. 2004. Knowledge Rep-
resentation and Reasoning. San Francisco, California:
Morgan Kaufmann.
Brassens, G. 1993. Pomes & Chansons. Paris: ditions
du Seuil.
Cambouropoulos, E., et al. 2005. A Pattern Extraction
Algorithm for Abstract Melodic Representations That
Allow Partial Overlapping of Intervallic Categories.
Proceedings of the Sixth Annual International Confer-
ence on Music Information Retrieval. London: Queen
Mary University of London, pp. 167174.
Conklin, D. 2002. Representation and Discovery of
Vertical Patterns in Music. In C. Anagnostopoulou,
M. Ferrand, and A. Smaill, eds. Music and Artificial
Intelligence: Lecture Notes in Artificial Intelligence
2445. Berlin: Springer-Verlag, pp. 3242.
Conklin, D. 2006. Melodic Analysis with Segment
Classes. Machine Learning 65:349360.
Conklin, D., and C. Anagnostopoulou. 2001. Representa-
tion and Discovery of Multiple Viewpoint Patterns.
Proceedings of the 2001 International Computer Music
Conference. San Francisco, California: International
Computer Music Association, pp. 479485.
Conklin, D., and C. Anagnostopoulou. 2005. Segmental
Pattern Discovery in Music. INFORMS Journal on
Computing 18(13):285293.
Conklin, D., and M. Bergeron. 2007. Discovery of Gen-
eralized Interval Patterns. Proceedings of the Fourth
Sound and Music Computing Conference. Lefkada,
Greece: National and Kapodistrian University of Ath-
ens, pp. 149152.
Conklin, D., and I. H. Witten. 1995. Multiple Viewpoint
Systems for Music Prediction. Journal of New Music
Research 24(1):5173.
Cope, D. 1991. Computers and Musical Style. Madison,
Wisconsin: A-R Editions.
Gusfield, D. 1997. Algorithms on Strings, Trees, and Se-
quences. Cambridge, UK: Cambridge University Press.
70
Huron, D. 2001. What is a Musical Feature? Fortes
Analysis of Brahmss Opus 51, No. 1, Revisited. Mu-
sic Theory Online 7(4). Available online at www
.societymusictheory.org / mto / .
Keefe, S. 2002. Sophisticated Simplicity: Text and Music
in the Early Songs of Georges Brassens. Tijdschrift
voor Muziektheorie 7(1):1123.
Lartillot, O. 2004. A Musical Pattern Discovery System
Founded on a Modeling of Listening Strategies. Com-
puter Music Journal 28(3):5367.
Lawrence, C., et al. 1993. Detecting Subtle Sequence
Signals: A Gibbs Sampling Strategy for Multiple Align-
ment. Science 262(5131):208214.
Lin, C.-R., et al. 2004. Music Classification Using Sig-
nificant Repeated Patterns. Proceedings of the Eighth
International Conference on Database Systems for
Advanced Applications. Berlin: Springer, pp. 508518.
Manning, C., and H. Schutze. 1999. Foundations of
Statistical Natural Language Processing. Cambridge,
Massachusetts: MIT Press.
Mongeau, M., and D. Sankoff. 1990. Comparison of
Musical Sequences. Computers and the Humanities
24(3):161175.
Rowe, R. 1993. Interactive Music Systems: Machine
Listening and Composing. Cambridge, Massachusetts:
MIT Press.
Ruwet, N. 1966. Mthodes dAnalyse en Musicologie.
Revue Belge de Musicologie 20:6590.
Schaffrath, H. 1995. The Essen Folksong Collection in
the Humdrum Kern Format. Stanford, California:
Center for Computer Assisted Research in the Human-
ities at Stanford University.
Uno, T., et al. 2003. LCM: An Efficient Algorithm
Enumerating Frequent Closed Item Sets. Proceed-
ings of the Workshop on Frequent Itemset Mining
Implementations. Aachen, Germany: CEUR Workshop
Proceedings.
Computer Music Journal